Efficient Alzheimer’s Stage Classification Using Swin Transformer from Paired MRI Planes

This study uses Swin Transformer (ST), a local, hierarchical attention mechanism with shifted windows, to effectively and computationally efficiently classify Alzheimer’s disease (AD) stages from 3D MRI data. MRI images from the ADNI dataset that represented the stages of advanced AD (AAD), mild cognitive impairment (MCI), and cognitively normal (CN) were considered for validation. The most informative 10–12 2D MRI slices from axial (A<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x</inf>), sagittal (S<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">g</inf>), and coronal (C<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</inf>) planes were selected for three experimental cases: A<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x_</inf> C<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</inf>, A<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x_</inf> S<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">g</inf>, and C<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r_</inf> S<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">g</inf>. The ST achieved comparable accuracy (A<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</inf>: 92.9%), with ST’s improved sensitivity (Sen: 94%) and specificity (Spe: 97%), while reduced computational complexity by 95% w.r.t to use of traditional Vision Transformers. Further, a pre-trained ST model achieved A<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</inf> upto 94.5%, demonstrating its potential for real-world MRI based analysis.